[X86][FP16] Fix vector_shuffle and lowering without f16c feature problems

[pengfei]

Summary

The problem Alexander reported on D127982 was caused by an optimization
for AVX512-FP16 instruction. We must limit it to the feature enabled only.

During the investigation, I found we didn't expand for fp_round/fp_extend
without F16C. This may result runtime crash, so change them too.

[RKSimon]

llvm/test/CodeGen/X86/vector-half-conversions.ll

8

please don't use COMMON when it isn't - tbh I'd prefer you just kept F16C + AVX512 seperate

[pengfei]

Split F16C and AVX512.

[LuoYuanke]

llvm/lib/Target/X86/X86ISelLowering.cpp

16083–16093

It seems the condition can be refined to hasAVX(), or we can add hasAVX() check in lowerShuffleAsBroadcast.
https://www.felixcloutier.com/x86/vbroadcast.html

[RKSimon]

llvm/test/CodeGen/X86/vector-half-conversions.ll

29

I'm confused as to under what cases we can't use vcvtph2ps for half -> float given all half values are exactly representable by float?

[pengfei]

llvm/lib/Target/X86/X86ISelLowering.cpp

16083–16093

No, we can't. We don't have a 16-bit broadcast until AVX512BW. We don't need extra work to leverage it since we have turn the shuffle to v8i16.

llvm/test/CodeGen/X86/vector-half-conversions.ll

29

Not sure if I understand you question correctly. vcvtph2ps requires F16C, which is independent of AVX and AVX2.

[RKSimon]

llvm/test/CodeGen/X86/vector-half-conversions.ll

29

Sorry - I missed the corresponding change in the attributes for the check prefixes.

It might make sense to pre-commit the changes to the RUN/check-prefixes so you can rebase the patch and show the codegen differences more clearly.

[pengfei]

llvm/test/CodeGen/X86/vector-half-conversions.ll

29

Yeah, I tried, but the test failed without the change.

[RKSimon]

LGTM - cheers

[alexfh]

@tstellar This commit needs to be cherrypicked to the 15.x branch, since it fixes a problem introduced before the llvm 16 cut (https://reviews.llvm.org/D127982).

[alexfh]

In D130817#3733569, @alexfh wrote:

@tstellar This commit needs to be cherrypicked to the 15.x branch, since it fixes a problem introduced before the llvm 16 cut (https://reviews.llvm.org/D127982).

Filed https://github.com/llvm/llvm-project/issues/57238

llvm/lib/Target/X86/X86ISelLowering.cpp

if (!Subtarget.useSoftFloat() && Subtarget.hasAVX()) {
     setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v8i16, MVT::v8i32);
     setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v8i16, MVT::v8i32);
     setOperationAction(ISD::FP_TO_SINT,                MVT::v8i32, Custom);
     setOperationAction(ISD::FP_TO_UINT,                MVT::v8i32, Custom);
     setOperationAction(ISD::STRICT_FP_TO_SINT,         MVT::v8i32, Custom);
 
     setOperationAction(ISD::SINT_TO_FP,         MVT::v8i32, Custom);
     setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v8i32, Custom);
+    setOperationAction(ISD::FP_EXTEND,          MVT::v8f32, Expand);
+    setOperationAction(ISD::FP_ROUND,           MVT::v8f16, Expand);
     setOperationAction(ISD::FP_EXTEND,          MVT::v4f64, Custom);
     setOperationAction(ISD::STRICT_FP_EXTEND,   MVT::v4f64, Custom);
 
     setOperationAction(ISD::STRICT_FP_ROUND,    MVT::v4f32, Legal);
     setOperationAction(ISD::STRICT_FADD,        MVT::v8f32, Legal);
     setOperationAction(ISD::STRICT_FADD,        MVT::v4f64, Legal);
     setOperationAction(ISD::STRICT_FSUB,        MVT::v8f32, Legal);
     setOperationAction(ISD::STRICT_FSUB,        MVT::v4f64, Legal);
static SDValue lowerV8F16Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
   assert(V1.getSimpleValueType() == MVT::v8f16 && "Bad operand type!");
   assert(V2.getSimpleValueType() == MVT::v8f16 && "Bad operand type!");
   assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
   int NumV2Elements = count_if(Mask, [](int M) { return M >= 8; });
 
+  if (Subtarget.hasFP16()) {
     if (NumV2Elements == 0) {
       // Check for being able to broadcast a single element.
       if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8f16, V1, V2,
                                                       Mask, Subtarget, DAG))
         return Broadcast;
     }
     if (NumV2Elements == 1 && Mask[0] >= 8)
-    if (SDValue V = lowerShuffleAsElementInsertion(DL, MVT::v8f16, V1, V2, Mask,
-                                                   Zeroable, Subtarget, DAG))
+      if (SDValue V = lowerShuffleAsElementInsertion(
+              DL, MVT::v8f16, V1, V2, Mask, Zeroable, Subtarget, DAG))
         return V;
+  }
 
   V1 = DAG.getBitcast(MVT::v8i16, V1);
   V2 = DAG.getBitcast(MVT::v8i16, V2);
   return DAG.getBitcast(MVT::v8f16,
                         DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, Mask));
 }
 
 // Lowers unary/binary shuffle as VPERMV/VPERMV3, for non-VLX targets,

llvm/test/CodeGen/X86/half.ll

 ; CHECK-I686-NEXT:  # %bb.1:
 ; CHECK-I686-NEXT:    movaps %xmm1, %xmm0
 ; CHECK-I686-NEXT:  .LBB23_2:
 ; CHECK-I686-NEXT:    retl
   %s = select i1 %c, <8 x half> %x, <8 x half> %y
   ret <8 x half> %s
 }
 
+define <8 x half> @shuffle(ptr %p) {
+; CHECK-LIBCALL-LABEL: shuffle:
+; CHECK-LIBCALL:       # %bb.0:
+; CHECK-LIBCALL-NEXT:    movdqu (%rdi), %xmm0
+; CHECK-LIBCALL-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
+; CHECK-LIBCALL-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,2,2,2]
+; CHECK-LIBCALL-NEXT:    retq
+;
+; BWON-F16C-LABEL: shuffle:
+; BWON-F16C:       # %bb.0:
+; BWON-F16C-NEXT:    vpshufhw {{.*#+}} xmm0 = mem[0,1,2,3,4,4,4,4]
+; BWON-F16C-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,2,2,2]
+; BWON-F16C-NEXT:    retq
+;
+; CHECK-I686-LABEL: shuffle:
+; CHECK-I686:       # %bb.0:
+; CHECK-I686-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; CHECK-I686-NEXT:    movdqu (%eax), %xmm0
+; CHECK-I686-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,4,4,4]
+; CHECK-I686-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,2,2,2]
+; CHECK-I686-NEXT:    retl
+  %1 = load <8 x half>, ptr %p, align 8
+  %2 = shufflevector <8 x half> %1, <8 x half> poison, <8 x i32> <i32 4, i32 4, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  ret <8 x half> %2
+}
+
 attributes #0 = { nounwind }

llvm/test/CodeGen/X86/vector-half-conversions.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx,+f16c -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx2,+f16c -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx2,+f16c,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx2,+f16c,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX512
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX512
-; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=ALL,AVX512
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx -verify-machineinstrs | FileCheck %s --check-prefixes=AVX,AVX1
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx2 -verify-machineinstrs | FileCheck %s --check-prefixes=AVX,AVX2
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+f16c -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,F16C
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+f16c -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,F16C
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+f16c,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,F16C
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+f16c,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,F16C
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,AVX512
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,AVX512
+; RUN: llc < %s -disable-peephole -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+fast-variable-perlane-shuffle -verify-machineinstrs | FileCheck %s --check-prefixes=COMMON,AVX512
 
 ;
 ; Half to Float
 ;
 
 define float @cvt_i16_to_f32(i16 %a0) nounwind {
-; ALL-LABEL: cvt_i16_to_f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    movzwl %di, %eax
-; ALL-NEXT:    vmovd %eax, %xmm0
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_i16_to_f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpinsrw $0, %edi, %xmm0, %xmm0
+; AVX-NEXT:    jmp __extendhfsf2@PLT # TAILCALL
+;
+; COMMON-LABEL: cvt_i16_to_f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    movzwl %di, %eax
+; COMMON-NEXT:    vmovd %eax, %xmm0
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast i16 %a0 to half
   %2 = fpext half %1 to float
   ret float %2
 }
 
 define <4 x float> @cvt_4i16_to_4f32(<4 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_4i16_to_4f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4i16_to_4f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movq %rax, %rdx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    # kill: def $eax killed $eax killed $rax
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrq $48, %rdx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4i16_to_4f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast <4 x i16> %a0 to <4 x half>
   %2 = fpext <4 x half> %1 to <4 x float>
   ret <4 x float> %2
 }
 
 define <4 x float> @cvt_8i16_to_4f32(<8 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_8i16_to_4f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8i16_to_4f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movq %rax, %rdx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    # kill: def $eax killed $eax killed $rax
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrq $48, %rdx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8i16_to_4f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = shufflevector <8 x i16> %a0, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %2 = bitcast <4 x i16> %1 to <4 x half>
   %3 = fpext <4 x half> %2 to <4 x float>
   ret <4 x float> %3
 }
 
 define <8 x float> @cvt_8i16_to_8f32(<8 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_8i16_to_8f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8i16_to_8f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $56, %rsp
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $56, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8i16_to_8f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = bitcast <8 x i16> %a0 to <8 x half>
   %2 = fpext <8 x half> %1 to <8 x float>
   ret <8 x float> %2
 }
 
 define <16 x float> @cvt_16i16_to_16f32(<16 x i16> %a0) nounwind {
-; AVX-LABEL: cvt_16i16_to_16f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtph2ps %xmm0, %ymm2
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX-NEXT:    vcvtph2ps %xmm0, %ymm1
-; AVX-NEXT:    vmovaps %ymm2, %ymm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: cvt_16i16_to_16f32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $104, %rsp
+; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    addq $104, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_16i16_to_16f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $104, %rsp
+; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    addq $104, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: cvt_16i16_to_16f32:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm2
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm1
+; F16C-NEXT:    vmovaps %ymm2, %ymm0
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: cvt_16i16_to_16f32:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps %ymm0, %zmm0
 ; AVX512-NEXT:    retq
   %1 = bitcast <16 x i16> %a0 to <16 x half>
   %2 = fpext <16 x half> %1 to <16 x float>
   ret <16 x float> %2
 }
 
 define <2 x float> @cvt_2i16_to_2f32_constrained(<2 x i16> %a0) nounwind strictfp {
-; ALL-LABEL: cvt_2i16_to_2f32_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_2i16_to_2f32_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $40, %rsp
+; AVX-NEXT:    vmovd %xmm0, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    addq $40, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_2i16_to_2f32_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast <2 x i16> %a0 to <2 x half>
   %2 = call <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <2 x float> %2
 }
 declare <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half>, metadata) strictfp
 
 define <4 x float> @cvt_4i16_to_4f32_constrained(<4 x i16> %a0) nounwind strictfp {
-; ALL-LABEL: cvt_4i16_to_4f32_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4i16_to_4f32_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movq %rax, %rdx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    # kill: def $eax killed $eax killed $rax
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrq $48, %rdx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4i16_to_4f32_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast <4 x i16> %a0 to <4 x half>
   %2 = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <4 x float> %2
 }
 declare <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half>, metadata) strictfp
 
 define <8 x float> @cvt_8i16_to_8f32_constrained(<8 x i16> %a0) nounwind strictfp {
-; ALL-LABEL: cvt_8i16_to_8f32_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8i16_to_8f32_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $56, %rsp
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $56, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8i16_to_8f32_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = bitcast <8 x i16> %a0 to <8 x half>
   %2 = call <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(<8 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <8 x float> %2
 }
 declare <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(<8 x half>, metadata) strictfp
 
 define <16 x float> @cvt_16i16_to_16f32_constrained(<16 x i16> %a0) nounwind strictfp {
-; AVX-LABEL: cvt_16i16_to_16f32_constrained:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX-NEXT:    vcvtph2ps %xmm1, %ymm1
-; AVX-NEXT:    vcvtph2ps %xmm0, %ymm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: cvt_16i16_to_16f32_constrained:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $104, %rsp
+; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    addq $104, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_16i16_to_16f32_constrained:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $104, %rsp
+; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovdqu {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $78, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vinsertps $16, (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $245, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    addq $104, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: cvt_16i16_to_16f32_constrained:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; F16C-NEXT:    vcvtph2ps %xmm1, %ymm1
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm0
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: cvt_16i16_to_16f32_constrained:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps %ymm0, %zmm0
 ; AVX512-NEXT:    retq
   %1 = bitcast <16 x i16> %a0 to <16 x half>
   %2 = call <16 x float> @llvm.experimental.constrained.fpext.v16f32.v16f16(<16 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <16 x float> %2
 }
 declare <16 x float> @llvm.experimental.constrained.fpext.v16f32.v16f16(<16 x half>, metadata) strictfp
 
 ;
 ; Half to Float (Load)
 ;
 
 define float @load_cvt_i16_to_f32(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_i16_to_f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    movzwl (%rdi), %eax
-; ALL-NEXT:    vmovd %eax, %xmm0
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_i16_to_f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    jmp __extendhfsf2@PLT # TAILCALL
+;
+; COMMON-LABEL: load_cvt_i16_to_f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    movzwl (%rdi), %eax
+; COMMON-NEXT:    vmovd %eax, %xmm0
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = load i16, ptr %a0
   %2 = bitcast i16 %1 to half
   %3 = fpext half %2 to float
   ret float %3
 }
 
 define <4 x float> @load_cvt_4i16_to_4f32(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_4i16_to_4f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_4i16_to_4f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vpinsrw $0, 6(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 2(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_4i16_to_4f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    retq
   %1 = load <4 x i16>, ptr %a0
   %2 = bitcast <4 x i16> %1 to <4 x half>
   %3 = fpext <4 x half> %2 to <4 x float>
   ret <4 x float> %3
 }
 
 define <4 x float> @load_cvt_8i16_to_4f32(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_8i16_to_4f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_8i16_to_4f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    movq (%rdi), %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movq %rax, %rdx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    # kill: def $eax killed $eax killed $rax
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrq $48, %rdx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_8i16_to_4f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %3 = bitcast <4 x i16> %2 to <4 x half>
   %4 = fpext <4 x half> %3 to <4 x float>
   ret <4 x float> %4
 }
 
 define <8 x float> @load_cvt_8i16_to_8f32(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_8i16_to_8f32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_8i16_to_8f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $48, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $48, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_8i16_to_8f32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %ymm0
+; COMMON-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = bitcast <8 x i16> %1 to <8 x half>
   %3 = fpext <8 x half> %2 to <8 x float>
   ret <8 x float> %3
 }
 
 define <16 x float> @load_cvt_16i16_to_16f32(ptr %a0) nounwind {
-; AVX-LABEL: load_cvt_16i16_to_16f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtph2ps (%rdi), %ymm0
-; AVX-NEXT:    vcvtph2ps 16(%rdi), %ymm1
-; AVX-NEXT:    retq
+; AVX1-LABEL: load_cvt_16i16_to_16f32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $80, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rbx), %xmm1
+; AVX1-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovaps 16(%rbx), %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 24(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 28(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 20(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    addq $80, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load_cvt_16i16_to_16f32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $80, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vpinsrw $0, 8(%rdi), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rbx), %xmm1
+; AVX2-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovaps 16(%rbx), %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm1[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 4(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 24(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 28(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 20(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    addq $80, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: load_cvt_16i16_to_16f32:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtph2ps (%rdi), %ymm0
+; F16C-NEXT:    vcvtph2ps 16(%rdi), %ymm1
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: load_cvt_16i16_to_16f32:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps (%rdi), %zmm0
 ; AVX512-NEXT:    retq
   %1 = load <16 x i16>, ptr %a0
   %2 = bitcast <16 x i16> %1 to <16 x half>
   %3 = fpext <16 x half> %2 to <16 x float>
   ret <16 x float> %3
 }
 
 define <4 x float> @load_cvt_4i16_to_4f32_constrained(ptr %a0) nounwind strictfp {
-; ALL-LABEL: load_cvt_4i16_to_4f32_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_4i16_to_4f32_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vpinsrw $0, 6(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 2(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vinsertps $16, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_4i16_to_4f32_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    retq
   %1 = load <4 x i16>, ptr %a0
   %2 = bitcast <4 x i16> %1 to <4 x half>
   %3 = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %2, metadata !"fpexcept.strict") strictfp
   ret <4 x float> %3
 }
 
 define <4 x float> @load_cvt_8i16_to_4f32_constrained(ptr %a0) nounwind strictfp {
-; ALL-LABEL: load_cvt_8i16_to_4f32_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_8i16_to_4f32_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    movq (%rdi), %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movq %rax, %rdx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    # kill: def $eax killed $eax killed $rax
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrq $48, %rdx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[2,3]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_8i16_to_4f32_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %3 = bitcast <4 x i16> %2 to <4 x half>
   %4 = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %3, metadata !"fpexcept.strict") strictfp
   ret <4 x float> %4
 }
 
 ;
 ; Half to Double
 ;
 
 define double @cvt_i16_to_f64(i16 %a0) nounwind {
-; ALL-LABEL: cvt_i16_to_f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    movzwl %di, %eax
-; ALL-NEXT:    vmovd %eax, %xmm0
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_i16_to_f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    vpinsrw $0, %edi, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    popq %rax
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_i16_to_f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    movzwl %di, %eax
+; COMMON-NEXT:    vmovd %eax, %xmm0
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast i16 %a0 to half
   %2 = fpext half %1 to double
   ret double %2
 }
 
 define <2 x double> @cvt_2i16_to_2f64(<2 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_2i16_to_2f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_2i16_to_2f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $40, %rsp
+; AVX-NEXT:    vmovd %xmm0, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    addq $40, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_2i16_to_2f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast <2 x i16> %a0 to <2 x half>
   %2 = fpext <2 x half> %1 to <2 x double>
   ret <2 x double> %2
 }
 
 define <4 x double> @cvt_4i16_to_4f64(<4 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_4i16_to_4f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4i16_to_4f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movl %eax, %edx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    shrq $48, %rax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrl $16, %edx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4i16_to_4f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = bitcast <4 x i16> %a0 to <4 x half>
   %2 = fpext <4 x half> %1 to <4 x double>
   ret <4 x double> %2
 }
 
 define <2 x double> @cvt_8i16_to_2f64(<8 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_8i16_to_2f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8i16_to_2f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $40, %rsp
+; AVX-NEXT:    vmovd %xmm0, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    addq $40, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8i16_to_2f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = shufflevector <8 x i16> %a0, <8 x i16> undef, <2 x i32> <i32 0, i32 1>
   %2 = bitcast <2 x i16> %1 to <2 x half>
   %3 = fpext <2 x half> %2 to <2 x double>
   ret <2 x double> %3
 }
 
 define <4 x double> @cvt_8i16_to_4f64(<8 x i16> %a0) nounwind {
-; ALL-LABEL: cvt_8i16_to_4f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8i16_to_4f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movl %eax, %edx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    shrq $48, %rax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrl $16, %edx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8i16_to_4f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = shufflevector <8 x i16> %a0, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %2 = bitcast <4 x i16> %1 to <4 x half>
   %3 = fpext <4 x half> %2 to <4 x double>
   ret <4 x double> %3
 }
 
 define <8 x double> @cvt_8i16_to_8f64(<8 x i16> %a0) nounwind {
 ; AVX-LABEL: cvt_8i16_to_8f64:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtph2ps %xmm0, %ymm1
-; AVX-NEXT:    vcvtps2pd %xmm1, %ymm0
-; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX-NEXT:    vcvtps2pd %xmm1, %ymm1
+; AVX-NEXT:    subq $88, %rsp
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $245, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $78, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT:    addq $88, %rsp
 ; AVX-NEXT:    retq
 ;
+; F16C-LABEL: cvt_8i16_to_8f64:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm1
+; F16C-NEXT:    vcvtps2pd %xmm1, %ymm0
+; F16C-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; F16C-NEXT:    vcvtps2pd %xmm1, %ymm1
+; F16C-NEXT:    retq
+;
 ; AVX512-LABEL: cvt_8i16_to_8f64:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %ymm0
 ; AVX512-NEXT:    vcvtps2pd %ymm0, %zmm0
 ; AVX512-NEXT:    retq
   %1 = bitcast <8 x i16> %a0 to <8 x half>
   %2 = fpext <8 x half> %1 to <8 x double>
   ret <8 x double> %2
 }
 
 define <2 x double> @cvt_2i16_to_2f64_constrained(<2 x i16> %a0) nounwind strictfp {
-; ALL-LABEL: cvt_2i16_to_2f64_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_2i16_to_2f64_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $40, %rsp
+; AVX-NEXT:    vmovd %xmm0, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    shrl $16, %eax
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm1
+; AVX-NEXT:    vmovdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    addq $40, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_2i16_to_2f64_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = bitcast <2 x i16> %a0 to <2 x half>
   %2 = call <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(<2 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <2 x double> %2
 }
 declare <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(<2 x half>, metadata) strictfp
 
 define <4 x double> @cvt_4i16_to_4f64_constrained(<4 x i16> %a0) nounwind strictfp {
-; ALL-LABEL: cvt_4i16_to_4f64_constrained:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4i16_to_4f64_constrained:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movl %eax, %edx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    shrq $48, %rax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrl $16, %edx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4i16_to_4f64_constrained:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = bitcast <4 x i16> %a0 to <4 x half>
   %2 = call <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(<4 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <4 x double> %2
 }
 declare <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(<4 x half>, metadata) strictfp
 
 define <8 x double> @cvt_8i16_to_8f64_constrained(<8 x i16> %a0) nounwind strictfp {
 ; AVX-LABEL: cvt_8i16_to_8f64_constrained:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtph2ps %xmm0, %ymm0
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX-NEXT:    vcvtps2pd %xmm1, %ymm1
-; AVX-NEXT:    vcvtps2pd %xmm0, %ymm0
+; AVX-NEXT:    subq $88, %rsp
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $245, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $78, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[2,3,0,1]
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT:    addq $88, %rsp
 ; AVX-NEXT:    retq
 ;
+; F16C-LABEL: cvt_8i16_to_8f64_constrained:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtph2ps %xmm0, %ymm0
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; F16C-NEXT:    vcvtps2pd %xmm1, %ymm1
+; F16C-NEXT:    vcvtps2pd %xmm0, %ymm0
+; F16C-NEXT:    retq
+;
 ; AVX512-LABEL: cvt_8i16_to_8f64_constrained:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %ymm0
 ; AVX512-NEXT:    vcvtps2pd %ymm0, %zmm0
 ; AVX512-NEXT:    retq
   %1 = bitcast <8 x i16> %a0 to <8 x half>
   %2 = call <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(<8 x half> %1, metadata !"fpexcept.strict") strictfp
   ret <8 x double> %2
 }
 declare <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(<8 x half>, metadata) strictfp
 
 ;
 ; Half to Double (Load)
 ;
 
 define double @load_cvt_i16_to_f64(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_i16_to_f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    movzwl (%rdi), %eax
-; ALL-NEXT:    vmovd %eax, %xmm0
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_i16_to_f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    popq %rax
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_i16_to_f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    movzwl (%rdi), %eax
+; COMMON-NEXT:    vmovd %eax, %xmm0
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = load i16, ptr %a0
   %2 = bitcast i16 %1 to half
   %3 = fpext half %2 to double
   ret double %3
 }
 
 define <2 x double> @load_cvt_2i16_to_2f64(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_2i16_to_2f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; ALL-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_2i16_to_2f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $40, %rsp
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 2(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    addq $40, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_2i16_to_2f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; COMMON-NEXT:    vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = load <2 x i16>, ptr %a0
   %2 = bitcast <2 x i16> %1 to <2 x half>
   %3 = fpext <2 x half> %2 to <2 x double>
   ret <2 x double> %3
 }
 
 define <4 x double> @load_cvt_4i16_to_4f64(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_4i16_to_4f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_4i16_to_4f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 2(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, 6(%rdi), %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_4i16_to_4f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = load <4 x i16>, ptr %a0
   %2 = bitcast <4 x i16> %1 to <4 x half>
   %3 = fpext <4 x half> %2 to <4 x double>
   ret <4 x double> %3
 }
 
 define <4 x double> @load_cvt_8i16_to_4f64(ptr %a0) nounwind {
-; ALL-LABEL: load_cvt_8i16_to_4f64:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtph2ps (%rdi), %xmm0
-; ALL-NEXT:    vcvtps2pd %xmm0, %ymm0
-; ALL-NEXT:    retq
+; AVX-LABEL: load_cvt_8i16_to_4f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    movq (%rdi), %rax
+; AVX-NEXT:    movq %rax, %rcx
+; AVX-NEXT:    movl %eax, %edx
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    shrq $48, %rax
+; AVX-NEXT:    shrq $32, %rcx
+; AVX-NEXT:    shrl $16, %edx
+; AVX-NEXT:    vpinsrw $0, %edx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %ecx, %xmm0, %xmm0
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vunpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: load_cvt_8i16_to_4f64:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtph2ps (%rdi), %xmm0
+; COMMON-NEXT:    vcvtps2pd %xmm0, %ymm0
+; COMMON-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = shufflevector <8 x i16> %1, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %3 = bitcast <4 x i16> %2 to <4 x half>
   %4 = fpext <4 x half> %3 to <4 x double>
   ret <4 x double> %4
 }
 
 define <8 x double> @load_cvt_8i16_to_8f64(ptr %a0) nounwind {
-; AVX-LABEL: load_cvt_8i16_to_8f64:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtph2ps (%rdi), %ymm1
-; AVX-NEXT:    vcvtps2pd %xmm1, %ymm0
-; AVX-NEXT:    vextractf128 $1, %ymm1, %xmm1
-; AVX-NEXT:    vcvtps2pd %xmm1, %ymm1
-; AVX-NEXT:    retq
+; AVX1-LABEL: load_cvt_8i16_to_8f64:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $80, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpinsrw $0, 8(%rbx), %xmm0, %xmm0
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX1-NEXT:    callq __extendhfsf2@PLT
+; AVX1-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX1-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    addq $80, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load_cvt_8i16_to_8f64:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $80, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 4(%rdi), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrlq $48, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 12(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpinsrw $0, 8(%rbx), %xmm0, %xmm0
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpsrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX2-NEXT:    callq __extendhfsf2@PLT
+; AVX2-NEXT:    vcvtss2sd %xmm0, %xmm0, %xmm0
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; AVX2-NEXT:    vinsertf128 $1, (%rsp), %ymm0, %ymm1 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    addq $80, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: load_cvt_8i16_to_8f64:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtph2ps (%rdi), %ymm1
+; F16C-NEXT:    vcvtps2pd %xmm1, %ymm0
+; F16C-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; F16C-NEXT:    vcvtps2pd %xmm1, %ymm1
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: load_cvt_8i16_to_8f64:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtph2ps (%rdi), %ymm0
 ; AVX512-NEXT:    vcvtps2pd %ymm0, %zmm0
 ; AVX512-NEXT:    retq
   %1 = load <8 x i16>, ptr %a0
   %2 = bitcast <8 x i16> %1 to <8 x half>
   %3 = fpext <8 x half> %2 to <8 x double>
   ret <8 x double> %3
 }
 
 ;
 ; Float to Half
 ;
 
 define i16 @cvt_f32_to_i16(float %a0) nounwind {
-; ALL-LABEL: cvt_f32_to_i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    vmovd %xmm0, %eax
-; ALL-NEXT:    # kill: def $ax killed $ax killed $eax
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_f32_to_i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, %eax
+; AVX-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX-NEXT:    popq %rcx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_f32_to_i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    vmovd %xmm0, %eax
+; COMMON-NEXT:    # kill: def $ax killed $ax killed $eax
+; COMMON-NEXT:    retq
   %1 = fptrunc float %a0 to half
   %2 = bitcast half %1 to i16
   ret i16 %2
 }
 
 define <4 x i16> @cvt_4f32_to_4i16(<4 x float> %a0) nounwind {
-; ALL-LABEL: cvt_4f32_to_4i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4f32_to_4i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpshufd $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4f32_to_4i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   ret <4 x i16> %2
 }
 
 define <8 x i16> @cvt_4f32_to_8i16_undef(<4 x float> %a0) nounwind {
-; ALL-LABEL: cvt_4f32_to_8i16_undef:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4f32_to_8i16_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpshufd $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4f32_to_8i16_undef:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   ret <8 x i16> %3
 }
 
 define <8 x i16> @cvt_4f32_to_8i16_zero(<4 x float> %a0) nounwind {
-; ALL-LABEL: cvt_4f32_to_8i16_zero:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_4f32_to_8i16_zero:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $72, %rsp
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpshufd $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX-NEXT:    addq $72, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4f32_to_8i16_zero:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   ret <8 x i16> %3
 }
 
 define <8 x i16> @cvt_8f32_to_8i16(<8 x float> %a0) nounwind {
-; ALL-LABEL: cvt_8f32_to_8i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    retq
+; AVX-LABEL: cvt_8f32_to_8i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    subq $88, %rsp
+; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    addq $88, %rsp
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_8f32_to_8i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    retq
   %1 = fptrunc <8 x float> %a0 to <8 x half>
   %2 = bitcast <8 x half> %1 to <8 x i16>
   ret <8 x i16> %2
 }
 
 define <16 x i16> @cvt_16f32_to_16i16(<16 x float> %a0) nounwind {
-; AVX-LABEL: cvt_16f32_to_16i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
-; AVX-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
-; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: cvt_16f32_to_16i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $120, %rsp
+; AVX1-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    addq $120, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_16f32_to_16i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $120, %rsp
+; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vinserti128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    addq $120, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: cvt_16f32_to_16i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; F16C-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; F16C-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: cvt_16f32_to_16i16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtps2ph $4, %zmm0, %ymm0
 ; AVX512-NEXT:    retq
   %1 = fptrunc <16 x float> %a0 to <16 x half>
   %2 = bitcast <16 x half> %1 to <16 x i16>
   ret <16 x i16> %2
 }
 
 ;
 ; Float to Half (Store)
 ;
 
 define void @store_cvt_f32_to_i16(float %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_f32_to_i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    vmovd %xmm0, %eax
-; ALL-NEXT:    movw %ax, (%rdi)
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_f32_to_i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_f32_to_i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    vmovd %xmm0, %eax
+; COMMON-NEXT:    movw %ax, (%rdi)
+; COMMON-NEXT:    retq
   %1 = fptrunc float %a0 to half
   %2 = bitcast half %1 to i16
   store i16 %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f32_to_4i16(<4 x float> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_4f32_to_4i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, (%rdi)
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_4f32_to_4i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $64, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpextrw $0, %xmm0, 6(%rbx)
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpextrw $0, %xmm0, 4(%rbx)
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpextrw $0, %xmm0, 2(%rbx)
+; AVX-NEXT:    addq $64, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_4f32_to_4i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, (%rdi)
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   store <4 x i16> %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f32_to_8i16_undef(<4 x float> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_4f32_to_8i16_undef:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    vmovaps %xmm0, (%rdi)
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_4f32_to_8i16_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $64, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpshufd $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX-NEXT:    addq $64, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_4f32_to_8i16_undef:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    vmovaps %xmm0, (%rdi)
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   store <8 x i16> %3, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f32_to_8i16_zero(<4 x float> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_4f32_to_8i16_zero:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; ALL-NEXT:    vmovaps %xmm0, (%rdi)
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_4f32_to_8i16_zero:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $64, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpshufd $255, (%rsp), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX-NEXT:    addq $64, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_4f32_to_8i16_zero:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
+; COMMON-NEXT:    vmovaps %xmm0, (%rdi)
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x float> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   store <8 x i16> %3, ptr %a1
   ret void
 }
 
 define void @store_cvt_8f32_to_8i16(<8 x float> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_8f32_to_8i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_8f32_to_8i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $80, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX-NEXT:    vzeroupper
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX-NEXT:    callq __truncsfhf2@PLT
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX-NEXT:    vmovdqa %xmm0, (%rbx)
+; AVX-NEXT:    addq $80, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_8f32_to_8i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    retq
   %1 = fptrunc <8 x float> %a0 to <8 x half>
   %2 = bitcast <8 x half> %1 to <8 x i16>
   store <8 x i16> %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_16f32_to_16i16(<16 x float> %a0, ptr %a1) nounwind {
-; AVX-LABEL: store_cvt_16f32_to_16i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtps2ph $4, %ymm1, 16(%rdi)
-; AVX-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; AVX1-LABEL: store_cvt_16f32_to_16i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $112, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovaps %ymm0, (%rbx)
+; AVX1-NEXT:    addq $112, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: store_cvt_16f32_to_16i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $112, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vinserti128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovdqa %ymm0, (%rbx)
+; AVX2-NEXT:    addq $112, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: store_cvt_16f32_to_16i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtps2ph $4, %ymm1, 16(%rdi)
+; F16C-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: store_cvt_16f32_to_16i16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtps2ph $4, %zmm0, (%rdi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %1 = fptrunc <16 x float> %a0 to <16 x half>
   %2 = bitcast <16 x half> %1 to <16 x i16>

 ; ALL-LABEL: cvt_f64_to_i16:
 ; ALL:       # %bb.0:
 ; ALL-NEXT:    pushq %rax
 ; ALL-NEXT:    callq __truncdfhf2@PLT
 ; ALL-NEXT:    vpextrw $0, %xmm0, %eax
 ; ALL-NEXT:    # kill: def $ax killed $ax killed $eax
 ; ALL-NEXT:    popq %rcx
 ; ALL-NEXT:    retq
+; AVX-LABEL: cvt_f64_to_i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rax
+; AVX-NEXT:    callq __truncdfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, %eax
+; AVX-NEXT:    # kill: def $ax killed $ax killed $eax
+; AVX-NEXT:    popq %rcx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: cvt_f64_to_i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    pushq %rax
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vpextrw $0, %xmm0, %eax
+; COMMON-NEXT:    # kill: def $ax killed $ax killed $eax
+; COMMON-NEXT:    popq %rcx
+; COMMON-NEXT:    retq
   %1 = fptrunc double %a0 to half
   %2 = bitcast half %1 to i16
   ret i16 %2
 }
 
 define <2 x i16> @cvt_2f64_to_2i16(<2 x double> %a0) nounwind {
 ; AVX-LABEL: cvt_2f64_to_2i16:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    subq $40, %rsp
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
 ; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
+; AVX-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
 ; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
 ; AVX-NEXT:    addq $40, %rsp
 ; AVX-NEXT:    retq
+;
+; F16C-LABEL: cvt_2f64_to_2i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    subq $40, %rsp
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    addq $40, %rsp
+; F16C-NEXT:    retq
   %1 = fptrunc <2 x double> %a0 to <2 x half>
   %2 = bitcast <2 x half> %1 to <2 x i16>
   ret <2 x i16> %2
 }
 
 define <4 x i16> @cvt_4f64_to_4i16(<4 x double> %a0) nounwind {
-; AVX-LABEL: cvt_4f64_to_4i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    subq $72, %rsp
-; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
-; AVX-NEXT:    addq $72, %rsp
-; AVX-NEXT:    retq
+; AVX1-LABEL: cvt_4f64_to_4i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $88, %rsp
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX1-NEXT:    addq $88, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_4f64_to_4i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $88, %rsp
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX2-NEXT:    addq $88, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: cvt_4f64_to_4i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    subq $72, %rsp
+; F16C-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
+; F16C-NEXT:    addq $72, %rsp
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: cvt_4f64_to_4i16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    subq $72, %rsp
 ; AVX512-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX512-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX512-NEXT:    vzeroupper

 ; AVX512-NEXT:    addq $72, %rsp
 ; AVX512-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   ret <4 x i16> %2
 }
 
 define <8 x i16> @cvt_4f64_to_8i16_undef(<4 x double> %a0) nounwind {
-; AVX-LABEL: cvt_4f64_to_8i16_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    subq $72, %rsp
-; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
-; AVX-NEXT:    addq $72, %rsp
-; AVX-NEXT:    retq
+; AVX1-LABEL: cvt_4f64_to_8i16_undef:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $88, %rsp
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX1-NEXT:    addq $88, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_4f64_to_8i16_undef:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $88, %rsp
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX2-NEXT:    addq $88, %rsp
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: cvt_4f64_to_8i16_undef:
+; F16C:       # %bb.0:
+; F16C-NEXT:    subq $72, %rsp
+; F16C-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
+; F16C-NEXT:    addq $72, %rsp
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: cvt_4f64_to_8i16_undef:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    subq $72, %rsp
 ; AVX512-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512-NEXT:    vextractf128 $1, %ymm0, %xmm0
 ; AVX512-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX512-NEXT:    vzeroupper

 ; AVX512-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   ret <8 x i16> %3
 }
 
 define <8 x i16> @cvt_4f64_to_8i16_zero(<4 x double> %a0) nounwind {
-; ALL-LABEL: cvt_4f64_to_8i16_zero:
-; ALL:       # %bb.0:
-; ALL-NEXT:    subq $72, %rsp
-; ALL-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; ALL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
-; ALL-NEXT:    addq $72, %rsp
-; ALL-NEXT:    retq
+; AVX1-LABEL: cvt_4f64_to_8i16_zero:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    subq $88, %rsp
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX1-NEXT:    addq $88, %rsp
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: cvt_4f64_to_8i16_zero:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    subq $88, %rsp
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX2-NEXT:    addq $88, %rsp
+; AVX2-NEXT:    retq
+;
+; COMMON-LABEL: cvt_4f64_to_8i16_zero:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    subq $72, %rsp
+; COMMON-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; COMMON-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; COMMON-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
+; COMMON-NEXT:    addq $72, %rsp
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   ret <8 x i16> %3
 }
 
 define <8 x i16> @cvt_8f64_to_8i16(<8 x double> %a0) nounwind {
 ; AVX-LABEL: cvt_8f64_to_8i16:

 ; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
 ; AVX-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
 ; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 ; AVX-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
 ; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
 ; AVX-NEXT:    addq $104, %rsp
 ; AVX-NEXT:    retq
 ;
+; F16C-LABEL: cvt_8f64_to_8i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    subq $104, %rsp
+; F16C-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; F16C-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0]
+; F16C-NEXT:    addq $104, %rsp
+; F16C-NEXT:    retq
+;
 ; AVX512-LABEL: cvt_8f64_to_8i16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    subq $120, %rsp
 ; AVX512-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-NEXT:    vextractf32x4 $3, %zmm0, %xmm0
 ; AVX512-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    callq __truncdfhf2@PLT

 ; ALL-LABEL: store_cvt_f64_to_i16:
 ; ALL:       # %bb.0:
 ; ALL-NEXT:    pushq %rbx
 ; ALL-NEXT:    movq %rdi, %rbx
 ; ALL-NEXT:    callq __truncdfhf2@PLT
 ; ALL-NEXT:    vpextrw $0, %xmm0, (%rbx)
 ; ALL-NEXT:    popq %rbx
 ; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_f64_to_i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    callq __truncdfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_f64_to_i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    pushq %rbx
+; COMMON-NEXT:    movq %rdi, %rbx
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; COMMON-NEXT:    popq %rbx
+; COMMON-NEXT:    retq
   %1 = fptrunc double %a0 to half
   %2 = bitcast half %1 to i16
   store i16 %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_2f64_to_2i16(<2 x double> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_2f64_to_2i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    pushq %rbx
-; ALL-NEXT:    subq $32, %rsp
-; ALL-NEXT:    movq %rdi, %rbx
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vmovd %xmm0, (%rbx)
-; ALL-NEXT:    addq $32, %rsp
-; ALL-NEXT:    popq %rbx
-; ALL-NEXT:    retq
+; AVX-LABEL: store_cvt_2f64_to_2i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $32, %rsp
+; AVX-NEXT:    movq %rdi, %rbx
+; AVX-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX-NEXT:    callq __truncdfhf2@PLT
+; AVX-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __truncdfhf2@PLT
+; AVX-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX-NEXT:    vpextrw $0, %xmm0, 2(%rbx)
+; AVX-NEXT:    addq $32, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_2f64_to_2i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    pushq %rbx
+; COMMON-NEXT:    subq $32, %rsp
+; COMMON-NEXT:    movq %rdi, %rbx
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vmovd %xmm0, (%rbx)
+; COMMON-NEXT:    addq $32, %rsp
+; COMMON-NEXT:    popq %rbx
+; COMMON-NEXT:    retq
   %1 = fptrunc <2 x double> %a0 to <2 x half>
   %2 = bitcast <2 x half> %1 to <2 x i16>
   store <2 x i16> %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f64_to_4i16(<4 x double> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_4f64_to_4i16:
-; ALL:       # %bb.0:
-; ALL-NEXT:    pushq %rbx
-; ALL-NEXT:    subq $64, %rsp
-; ALL-NEXT:    movq %rdi, %rbx
-; ALL-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; ALL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; ALL-NEXT:    vmovq %xmm0, (%rbx)
-; ALL-NEXT:    addq $64, %rsp
-; ALL-NEXT:    popq %rbx
-; ALL-NEXT:    retq
+; AVX1-LABEL: store_cvt_4f64_to_4i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $80, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpextrw $0, %xmm0, 4(%rbx)
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpextrw $0, %xmm0, 6(%rbx)
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    vpextrw $0, %xmm0, 2(%rbx)
+; AVX1-NEXT:    addq $80, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: store_cvt_4f64_to_4i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $80, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpextrw $0, %xmm0, 4(%rbx)
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpextrw $0, %xmm0, (%rbx)
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpextrw $0, %xmm0, 6(%rbx)
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    vpextrw $0, %xmm0, 2(%rbx)
+; AVX2-NEXT:    addq $80, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_4f64_to_4i16:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    pushq %rbx
+; COMMON-NEXT:    subq $64, %rsp
+; COMMON-NEXT:    movq %rdi, %rbx
+; COMMON-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; COMMON-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; COMMON-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; COMMON-NEXT:    vmovq %xmm0, (%rbx)
+; COMMON-NEXT:    addq $64, %rsp
+; COMMON-NEXT:    popq %rbx
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   store <4 x i16> %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f64_to_8i16_undef(<4 x double> %a0, ptr %a1) nounwind {
-; AVX-LABEL: store_cvt_4f64_to_8i16_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    pushq %rbx
-; AVX-NEXT:    subq $64, %rsp
-; AVX-NEXT:    movq %rdi, %rbx
-; AVX-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; AVX-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; AVX-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = mem[1,0]
-; AVX-NEXT:    callq __truncdfhf2@PLT
-; AVX-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; AVX-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; AVX-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
-; AVX-NEXT:    vmovaps %xmm0, (%rbx)
-; AVX-NEXT:    addq $64, %rsp
-; AVX-NEXT:    popq %rbx
-; AVX-NEXT:    retq
+; AVX1-LABEL: store_cvt_4f64_to_8i16_undef:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $80, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX1-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX1-NEXT:    addq $80, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: store_cvt_4f64_to_8i16_undef:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $80, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX2-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX2-NEXT:    addq $80, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: store_cvt_4f64_to_8i16_undef:
+; F16C:       # %bb.0:
+; F16C-NEXT:    pushq %rbx
+; F16C-NEXT:    subq $64, %rsp
+; F16C-NEXT:    movq %rdi, %rbx
+; F16C-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
+; F16C-NEXT:    vmovaps %xmm0, (%rbx)
+; F16C-NEXT:    addq $64, %rsp
+; F16C-NEXT:    popq %rbx
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: store_cvt_4f64_to_8i16_undef:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    pushq %rbx
 ; AVX512-NEXT:    subq $64, %rsp
 ; AVX512-NEXT:    movq %rdi, %rbx
 ; AVX512-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
 ; AVX512-NEXT:    vextractf128 $1, %ymm0, %xmm0
; AVX512-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   store <8 x i16> %3, ptr %a1
   ret void
 }
 
 define void @store_cvt_4f64_to_8i16_zero(<4 x double> %a0, ptr %a1) nounwind {
-; ALL-LABEL: store_cvt_4f64_to_8i16_zero:
-; ALL:       # %bb.0:
-; ALL-NEXT:    pushq %rbx
-; ALL-NEXT:    subq $64, %rsp
-; ALL-NEXT:    movq %rdi, %rbx
-; ALL-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
-; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; ALL-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
-; ALL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; ALL-NEXT:    vzeroupper
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
-; ALL-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = mem[1,0]
-; ALL-NEXT:    callq __truncdfhf2@PLT
-; ALL-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
-; ALL-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
-; ALL-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
-; ALL-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
-; ALL-NEXT:    vmovaps %xmm0, (%rbx)
-; ALL-NEXT:    addq $64, %rsp
-; ALL-NEXT:    popq %rbx
-; ALL-NEXT:    retq
+; AVX1-LABEL: store_cvt_4f64_to_8i16_zero:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $80, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncdfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX1-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX1-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX1-NEXT:    addq $80, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: store_cvt_4f64_to_8i16_zero:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $80, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovupd %ymm0, (%rsp) # 32-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovupd (%rsp), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovapd %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vmovapd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncdfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1, %xmm1 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1],xmm1[2],mem[2],xmm1[3],mem[3]
+; AVX2-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0],xmm0[0],zero,zero
+; AVX2-NEXT:    vmovaps %xmm0, (%rbx)
+; AVX2-NEXT:    addq $80, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    retq
+;
+; COMMON-LABEL: store_cvt_4f64_to_8i16_zero:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    pushq %rbx
+; COMMON-NEXT:    subq $64, %rsp
+; COMMON-NEXT:    movq %rdi, %rbx
+; COMMON-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; COMMON-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; COMMON-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; COMMON-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; COMMON-NEXT:    vzeroupper
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; COMMON-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = mem[1,0]
+; COMMON-NEXT:    callq __truncdfhf2@PLT
+; COMMON-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; COMMON-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; COMMON-NEXT:    vinsertps $28, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; COMMON-NEXT:    # xmm0 = xmm0[0],mem[0],zero,zero
+; COMMON-NEXT:    vmovaps %xmm0, (%rbx)
+; COMMON-NEXT:    addq $64, %rsp
+; COMMON-NEXT:    popq %rbx
+; COMMON-NEXT:    retq
   %1 = fptrunc <4 x double> %a0 to <4 x half>
   %2 = bitcast <4 x half> %1 to <4 x i16>
   %3 = shufflevector <4 x i16> %2, <4 x i16> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   store <8 x i16> %3, ptr %a1
   ret void
 }
 
 define void @store_cvt_8f64_to_8i16(<8 x double> %a0, ptr %a1) nounwind {

 ; AVX-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 ; AVX-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
 ; AVX-NEXT:    # xmm0 = xmm0[0],mem[0]
 ; AVX-NEXT:    vmovdqa %xmm0, (%rbx)
 ; AVX-NEXT:    addq $96, %rsp
 ; AVX-NEXT:    popq %rbx
 ; AVX-NEXT:    retq
 ;
+; F16C-LABEL: store_cvt_8f64_to_8i16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    pushq %rbx
+; F16C-NEXT:    subq $96, %rsp
+; F16C-NEXT:    movq %rdi, %rbx
+; F16C-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; F16C-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, (%rsp), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; F16C-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; F16C-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; F16C-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = mem[1,0]
+; F16C-NEXT:    callq __truncdfhf2@PLT
+; F16C-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; F16C-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; F16C-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; F16C-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; F16C-NEXT:    # xmm0 = xmm0[0],mem[0]
+; F16C-NEXT:    vmovdqa %xmm0, (%rbx)
+; F16C-NEXT:    addq $96, %rsp
+; F16C-NEXT:    popq %rbx
+; F16C-NEXT:    retq
+;
 ; AVX512-LABEL: store_cvt_8f64_to_8i16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    pushq %rbx
 ; AVX512-NEXT:    subq $112, %rsp
 ; AVX512-NEXT:    movq %rdi, %rbx
 ; AVX512-NEXT:    vmovups %zmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 64-byte Spill
 ; AVX512-NEXT:    vextractf32x4 $3, %zmm0, %xmm0
 ; AVX512-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill

 ; AVX512-NEXT:    retq
   %1 = fptrunc <8 x double> %a0 to <8 x half>
   %2 = bitcast <8 x half> %1 to <8 x i16>
   store <8 x i16> %2, ptr %a1
   ret void
 }
 
 define void @store_cvt_32f32_to_32f16(<32 x float> %a0, ptr %a1) nounwind {
-; AVX-LABEL: store_cvt_32f32_to_32f16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtps2ph $4, %ymm3, 48(%rdi)
-; AVX-NEXT:    vcvtps2ph $4, %ymm2, 32(%rdi)
-; AVX-NEXT:    vcvtps2ph $4, %ymm1, 16(%rdi)
-; AVX-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
-; AVX-NEXT:    vzeroupper
-; AVX-NEXT:    retq
+; AVX1-LABEL: store_cvt_32f32_to_32f16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    pushq %rbx
+; AVX1-NEXT:    subq $176, %rsp
+; AVX1-NEXT:    movq %rdi, %rbx
+; AVX1-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,0]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX1-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX1-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX1-NEXT:    callq __truncsfhf2@PLT
+; AVX1-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX1-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX1-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX1-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX1-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX1-NEXT:    vinsertf128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX1-NEXT:    vmovaps %ymm0, 32(%rbx)
+; AVX1-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX1-NEXT:    vmovaps %ymm0, (%rbx)
+; AVX1-NEXT:    addq $176, %rsp
+; AVX1-NEXT:    popq %rbx
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: store_cvt_32f32_to_32f16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    pushq %rbx
+; AVX2-NEXT:    subq $176, %rsp
+; AVX2-NEXT:    movq %rdi, %rbx
+; AVX2-NEXT:    vmovups %ymm3, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups %ymm2, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups %ymm1, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vextractf128 $1, %ymm1, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq (%rsp), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vinserti128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovdqu %ymm0, {{[-0-9]+}}(%r{{[sb]}}p) # 32-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,3,3,3]
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilps $255, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[3,3,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vpermilpd $1, {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,0]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vpunpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
+; AVX2-NEXT:    vmovdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovaps %xmm0, (%rsp) # 16-byte Spill
+; AVX2-NEXT:    vmovshdup {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = mem[1,1,3,3]
+; AVX2-NEXT:    callq __truncsfhf2@PLT
+; AVX2-NEXT:    vmovdqa (%rsp), %xmm1 # 16-byte Reload
+; AVX2-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; AVX2-NEXT:    vpunpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
+; AVX2-NEXT:    vpunpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 # 16-byte Folded Reload
+; AVX2-NEXT:    # xmm0 = xmm0[0],mem[0]
+; AVX2-NEXT:    vinserti128 $1, {{[-0-9]+}}(%r{{[sb]}}p), %ymm0, %ymm0 # 16-byte Folded Reload
+; AVX2-NEXT:    vmovdqa %ymm0, 32(%rbx)
+; AVX2-NEXT:    vmovups {{[-0-9]+}}(%r{{[sb]}}p), %ymm0 # 32-byte Reload
+; AVX2-NEXT:    vmovaps %ymm0, (%rbx)
+; AVX2-NEXT:    addq $176, %rsp
+; AVX2-NEXT:    popq %rbx
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; F16C-LABEL: store_cvt_32f32_to_32f16:
+; F16C:       # %bb.0:
+; F16C-NEXT:    vcvtps2ph $4, %ymm3, 48(%rdi)
+; F16C-NEXT:    vcvtps2ph $4, %ymm2, 32(%rdi)
+; F16C-NEXT:    vcvtps2ph $4, %ymm1, 16(%rdi)
+; F16C-NEXT:    vcvtps2ph $4, %ymm0, (%rdi)
+; F16C-NEXT:    vzeroupper
+; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: store_cvt_32f32_to_32f16:
 ; AVX512:       # %bb.0:
 ; AVX512-NEXT:    vcvtps2ph $4, %zmm1, 32(%rdi)
 ; AVX512-NEXT:    vcvtps2ph $4, %zmm0, (%rdi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %1 = fptrunc <32 x float> %a0 to <32 x half>
   store <32 x half> %1, ptr %a1
   ret void
 }
 
 define <4 x i32> @fptosi_2f16_to_4i32(<2 x half> %a) nounwind {
-; ALL-LABEL: fptosi_2f16_to_4i32:
-; ALL:       # %bb.0:
-; ALL-NEXT:    vpsrld $16, %xmm0, %xmm1
-; ALL-NEXT:    vpextrw $0, %xmm1, %eax
-; ALL-NEXT:    movzwl %ax, %eax
-; ALL-NEXT:    vmovd %eax, %xmm1
-; ALL-NEXT:    vcvtph2ps %xmm1, %xmm1
-; ALL-NEXT:    vcvttss2si %xmm1, %eax
-; ALL-NEXT:    vpextrw $0, %xmm0, %ecx
-; ALL-NEXT:    movzwl %cx, %ecx
-; ALL-NEXT:    vmovd %ecx, %xmm0
-; ALL-NEXT:    vcvtph2ps %xmm0, %xmm0
-; ALL-NEXT:    vcvttss2si %xmm0, %ecx
-; ALL-NEXT:    vmovd %ecx, %xmm0
-; ALL-NEXT:    vmovd %eax, %xmm1
-; ALL-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; ALL-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
-; ALL-NEXT:    retq
+; AVX-LABEL: fptosi_2f16_to_4i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    pushq %rbx
+; AVX-NEXT:    subq $16, %rsp
+; AVX-NEXT:    vmovdqa %xmm0, (%rsp) # 16-byte Spill
+; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvttss2si %xmm0, %ebx
+; AVX-NEXT:    vmovaps (%rsp), %xmm0 # 16-byte Reload
+; AVX-NEXT:    callq __extendhfsf2@PLT
+; AVX-NEXT:    vcvttss2si %xmm0, %eax
+; AVX-NEXT:    vmovd %eax, %xmm0
+; AVX-NEXT:    vmovd %ebx, %xmm1
+; AVX-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; AVX-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; AVX-NEXT:    addq $16, %rsp
+; AVX-NEXT:    popq %rbx
+; AVX-NEXT:    retq
+;
+; COMMON-LABEL: fptosi_2f16_to_4i32:
+; COMMON:       # %bb.0:
+; COMMON-NEXT:    vpsrld $16, %xmm0, %xmm1
+; COMMON-NEXT:    vpextrw $0, %xmm1, %eax
+; COMMON-NEXT:    movzwl %ax, %eax
+; COMMON-NEXT:    vmovd %eax, %xmm1
+; COMMON-NEXT:    vcvtph2ps %xmm1, %xmm1
+; COMMON-NEXT:    vcvttss2si %xmm1, %eax
+; COMMON-NEXT:    vpextrw $0, %xmm0, %ecx
+; COMMON-NEXT:    movzwl %cx, %ecx
+; COMMON-NEXT:    vmovd %ecx, %xmm0
+; COMMON-NEXT:    vcvtph2ps %xmm0, %xmm0
+; COMMON-NEXT:    vcvttss2si %xmm0, %ecx
+; COMMON-NEXT:    vmovd %ecx, %xmm0
+; COMMON-NEXT:    vmovd %eax, %xmm1
+; COMMON-NEXT:    vpunpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; COMMON-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; COMMON-NEXT:    retq
   %cvt = fptosi <2 x half> %a to <2 x i32>
   %ext = shufflevector <2 x i32> %cvt, <2 x i32> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x i32> %ext
 }